z.ai GLM under TAUSIK (Claude Code & Kilo)

TAUSIK runs on z.ai GLM models under any Anthropic-compatible host. GLM is a model family (axis-2, Decision #119) — pure data in model_profiles, not code — so it is independent of which host you run in. The simplest and most capable path is Claude Code: the host is unchanged, so every SENAR gate keeps firing and only the model field reads glm-*.

• Host (axis-1) = Claude Code / Kilo / Cursor / Qwen — owns the bootstrap directory, the MCP config, and active-model detection.
• Model family (axis-2) = Claude vs z.ai GLM — pure data. Switching or adding GLM models needs no code change.

z.ai's endpoint is Anthropic-compatible, so the session transcript looks exactly like Claude's — routing, verdicts, and cost all work unchanged.

Subscription, not per-token. The z.ai GLM Coding Plan (from ~$10/mo) is a flat-fee subscription with usage quotas — not pay-as-you-go API billing. Same spirit as running Claude Code on a Max/Pro plan; you keep working on a subscription rather than metered tokens.

---

1. GLM under Claude Code (recommended — subscription, full gates)

Keep Claude Code as your host and point it at z.ai. Because the host never changes, all SENAR enforcement gates keep firing (QG-0 no-code-without-task, QG-2 verify, scope / secret / firewall) — GLM simply becomes the brain.

Set two environment variables (shell profile or the IDE secret store — never commit them):

export ANTHROPIC_BASE_URL="https://api.z.ai/api/anthropic"
export ANTHROPIC_AUTH_TOKEN="<your-z.ai-key>"   # z.ai GLM Coding Plan key

Launch Claude Code. It now reasons through GLM on your z.ai subscription; TAUSIK reads model: glm-* from the transcript and routes within the GLM family (model_profiles, family glm). To recommend GLM tiers from the first message, set "default_family": "glm" in .tausik/config.json (see §4).

Secret hygiene: the z.ai key is a credential. Keep it in your shell profile or the IDE's secret store — never in .tausik/config.json, .kilo/, or anything tracked by git.

Smoke-test billing before you rely on it. z.ai sells the Coding Plan against its /api/coding/paas/v4 endpoint; confirm your Coding-Plan quota actually bills through the Anthropic-compatible /api/anthropic endpoint used above (send one request, check the z.ai dashboard) so usage draws on the subscription and not a pay-as-you-go wallet. z.ai documents the Coding Plan for Claude Code, but pin this down first.

The same two env vars also drive Kilo and any other Anthropic-compatible host — the sections below cover Kilo-specific bootstrap. Available GLM models today include glm-4.5-air, glm-4.6, and the glm-5.x line (§4).

2. Bootstrap TAUSIK for Kilo

python .tausik-lib/bootstrap/bootstrap.py --ide kilo

This writes the TAUSIK MCP server stanza to both known Kilo config paths (robust across Kilo versions — Decision #120):

• .kilo/kilo.jsonc (current kilo.ai docs)
• .kilocode/mcp.json (older Cline-lineage builds)

Both contain the same mcp entry:

{
  "mcp": {
    "tausik-project": {
      "type": "local",
      "command": ["<python>", "${workspaceFolder}/.kilo/mcp/project/server.py", "--project", "${workspaceFolder}"],
      "enabled": true
    }
  }
}

Paths are rename-proof: a server inside the project and --project use ${workspaceFolder} (Kilo expands it at launch), so renaming the project folder does not break the config. An external lib server keeps its absolute path. Existing servers and other keys are merged, not overwritten. Re-running is idempotent.

Restart Kilo after bootstrap so it loads the new MCP config.

If your Kilo build reads neither default path

Override the target(s) in .tausik/config.json:

{ "kilo": { "config_paths": ["kilo.jsonc"] } }

(paths are project-relative; the list fully replaces the defaults.)

3. Tell TAUSIK which GLM model is active

Kilo has no Claude-style JSONL transcript, so TAUSIK reads the active model from (in order):

1. the KILO_MODEL environment variable — e.g. export KILO_MODEL=glm-4.6
2. a model field in .kilo/kilo.json (or ~/.config/kilo/kilo.json)

With that set, task start shows GLM recommendations and correct under/over-powered verdicts. Without it, recommendations fall back to model_profiles.default_family (below) and then to Claude.

4. Switch / add GLM models — no code change

Defaults shipped in scripts/model_profiles.py:

Capability rank	GLM model
light (haiku)	glm-4.5-air
mid (sonnet)	glm-4.6
strong (opus)	glm-4.6
flagship (fable)	glm-4.6

Override or extend any of these — and pin GLM as the default family — in .tausik/config.json:

{
  "model_profiles": {
    "default_family": "glm",
    "families": {
      "glm": {
        "opus":  { "model": "glm-5.2", "display": "GLM-5.2" },
        "fable": { "model": "glm-5.2", "display": "GLM-5.2" }
      }
    }
  }
}

default_family: "glm" makes task start recommend GLM models even before any transcript/KILO_MODEL detection — ideal when you only ever run Kilo + z.ai.

How it fits together

Kilo Code (addon/CLI)  ──MCP──▶  tausik-project server  (.kilo/kilo.jsonc | .kilocode/mcp.json)
        │
        └── model: glm-4.6  ──▶  model_profiles (family=glm) ──▶ routing rank → glm model + verdict

The runtime is Kilo; the model is GLM. Neither knows about the other — that separation is what makes "TAUSIK in Kilo with any z.ai model" a config exercise, not a code one.